Seed Counting Method Based On Deep Learning And Computer Visio

Commercial and scientific plant breeding programs require phenotypic analysis of a large number of populations in which 1000-grain weight is an important measure of seed quality and prediction of yield.The efficient and accurate automatic counting method of crop seeds is beneficial to the development of modern agricultural production.Breeding innovation is closely related to the agricultural development model and is of great significance to food security and ecological security.The traditional method of 1000-grain weight measurement is time-consuming and laborious,and the accuracy can not be guaranteed because it depends on manual counting.Although the photoelectric counting method reduces the amount of labor,it has the problems of complex operation,high cost and low efficiency.The use of computer vision technology is a potential solution for some of these specific tasks.deep learning,especially convolutional neural networks(CNN),has shown many advantages over traditional methods in the past few years.in this work,we integrate computer vision and deep learning techniques to propose a vision-based seed counting method.The method is effective,simple in operation,low in cost and high in reliability.It can meet the requirement of fast and accurate counting of a large number of seeds in modern agricultural production.In order to realize the seed count based on vision,we need to collect the seed picture first.The seeds are placed on a white uniform board,the bottom is illuminated by a light source,and the top is photographed with a digital camera to obtain a clear and clean picture of the seeds.Next,the traditional computer vision method is used to preliminarily process the collected photos.Hough transform algorithm and water filling algorithm are used to automatically treat the clipping of detection area and correct it by morphological transformation.Finally,the expansion and corrosion algorithm is used to reduce the adhesion between seeds,and then the watershed algorithm is used to fill the seed edge,the initial segmentation of the image is processed by morphology and the connected region is marked to realize the seed count.It is found that the method is effective for regular shape seeds such as soybean segmentation,irregular shape and very small grain seeds such as rape and corn,there is a problem that the adhesion region can not be divided.In order to further solve the problem of poor segmentation of adhesion seed area,this paper analyzes the types of adhesion area,uses statistical method to determine the threshold of adhesion area,and cuts the adhesion area.Based on the deep learning theory,a seed classification and adhesion seed counting model based on convolution neural network is established to train the adhesion region into the network.The type of seed is determined by classification algorithm and the corresponding classification model is selected.Then the adhesion region is input into the convolution neural network to predict the actual number of seeds and complete the seed count.In this paper,8 kinds of seeds of soybean,red bean,rice,sweet corn,wheat,rape and corn were used as experimental objects to count the random number of seed images.In this paper,an image-based seed counting method is proposed,which is simple and easy to carry out.The automatic classification and counting can be realized only by placing the seeds on a uniform plate.The experimental results show that the method can quickly and effectively calculate the number of seeds in the picture,and has a certain universality.After adding the deep learning model,it also has a very good effect on the seeds with high adhesion degree,and has important practical significance for agricultural production such as harvest and breeding of 1000-grain heavy seeds.